Real teams do not standardize on one AI coding agent. One developer lives in Claude Code, another in Cursor, a third runs Codex from the terminal. They share a codebase, so their agents need to share context, without forcing everyone onto the same tool.
Tool choice has become personal. Some developers want a terminal-native agent, some want it inside their editor, some want a specific model. That diversity is healthy, but it creates a coordination problem: each agent has its own context window, its own session, and no idea the others exist. When three different agents edit the same area of the codebase in the same afternoon, the result is the same as if three people did it blindfolded.
Why the answer has to be agent-agnostic
The obvious fix, "everyone use the same agent," does not survive contact with a real team. People have preferences, contracts, and habits. A coordination layer that only works if everyone adopts one vendor is a non-starter. The layer has to be agent-agnostic: it must work the same whether a teammate is on Claude Code, Cursor, or Codex.
The common denominator that makes this possible is the Model Context Protocol (MCP). Claude Code, Cursor, and Codex all speak MCP. That means a single MCP server can expose the same coordination tools to all three, and each agent uses them natively, no per-tool plugin, no lowest-common-denominator hacks.
The mechanism: one shared room, three agents
The idea is a shared room that all of a team's agents connect to over MCP. The room holds three things, and nothing more: each agent's current intent, the active interface contracts, and alerts when a contract one agent depends on changes. No source code, no files, no diffs leave any machine. Only the small slice of interface metadata an agent explicitly publishes.
With Aethereum, each agent connects to the room with one command and then uses the same three tools regardless of which agent it is:
$ npx aethereum init # run once per machine, in each repo# afterwards, every agent has these MCP tools:# share_intent(text)# declare_contract(name, shape, dependsOn?)# get_team_context()A developer on Cursor changes the auth endpoint and their agent declares the new contract. A developer on Claude Code, on a different laptop, starts a session; their agent calls get_team_context() and immediately learns the auth contract changed and that it depends on it. A third developer on Codex shares intent that they are refactoring the same module, and the others see it. Three different agents, one coherent picture.
Setting it up per agent
Because the coordination tools are plain MCP, setup is the same idea in each agent, with the same outcome. aethereum init writes the right configuration for the agent it detects, including the hooks that let supported agents inject team context at session start and surface alerts before each prompt automatically. The documentation has the exact per-agent steps for Claude Code, Cursor, and Codex.
Auto mode derives a one-line summary of what an agent is working on and shares it as intent. It never sends your full prompt and never sends code. It is optional and can be turned off with AETHEREUM_AUTO=off. See the privacy policy for exactly what is and is not stored.
Why not just rely on Git or Slack?
Git only sees committed code, and the dangerous window is before anyone commits. Slack announcements rely on a human to write them and another human's agent to somehow know about them; no agent reads your channels. The whole point is to put the change where the agents already are: in their tool surface, as live context, the moment it happens. That is what closes the gap that causes 27.67% of AI agent pull requests to hit merge conflicts.
Takeaway
Keeping Claude Code, Cursor, and Codex in sync does not mean making everyone use the same agent. It means giving every agent, whatever it is, the same shared room over MCP, so a contract change made in one tool reaches every other tool as live context across machines. The team keeps its tool diversity, and the agents stop colliding. Get started with npx aethereum init; it is free to start.